The 7th International Telecommunications Symposium (ITS 2010)

Mobile Station Location using Genetic Algorithm Optimized Radio Frequency

Fingerprinting

Rafael Saraiva Campos and Lisandro Lovisolo

Faculdade de Engenharia
Universidade do Estado do Rio de Janeiro
Rio de Janeiro, Brazil
rafael saraiva@ig.com.br,lisandro@uerj.br

Abstract— Radio Frequency Fingerprinting estimates the
mobile station location by comparing a measured radio
frequency fingerprint with a set of previously collected or
generated reference fingerprints. This set is referred to as the
search or correlation space. Genetic algorithms can be used
to optimize both the location accuracy and the time required
to produce a position fix, reducing the size of the search
space. This paper proposes an innovation in such application of
genetic algorithms, restricting the first generation population to
the predicted best server area of the serving sector measured by
the mobile station. In field tests in a GSM cellular network in a
dense urban environment, this approach achieved reductions of
20% and 15% in the 50-th and 90-th percentile location errors,
respectively, in comparison to the original formulation, where
the initial population is randomly distributed throughout the
whole service area. An average reduction of 91% in the time
to produce a position fix was also observed.
Keywords— Mobile Station Location, Radio Frequency Fin-
gerprints, Propagation Modeling, Genetic Algorithms, First Gen-
eration Population, Correlation Space.
I. I NTRODUCTION
Different levels of location awareness are inherent to any
metropolitan area network (MAN) with wireless radio access.
These allow packet routing to any mobile station (MS) and
session continuity when the MS moves from one base transmission
station (BTS) coverage area to another. These built-in location
capabilities can be extended to provide added value location
services (LCS) to the subscribers, which include, among others,
location based billing and emergency call location. This latter
application has been receiving much attention from government
authorities and is already mandatory for cellular networks in some
countries [1] [2].
It seems that, in a near future, the winner technology for MS
positioning will be based on the Global Positioning System (GPS),
due to its accuracy and to the cost reduction in GPS receivers
as its deployment scales up. Nonetheless, even in such scenario,
there are still situations in which network based methods should
be employed: i) unavailability of GPS signals (such as in indoor
environments) and ii) location of MS without built-in GPS re-
ceivers. In addition, these methods serve as fallback when more
precise ones can not be used due to systemic limitations, increasing
LCS availability [3]. This latter improvement is of the highest
importance when considering the positioning of MS originating
emergency calls in cellular networks [4]. Network based methods
use information already available at the serving BTS or periodically
sent by the MS through the network measurement reports (NMR).
The NMRs are used for call or session management - power
control, handover, etc.
The location method presented in this work is a radio fre-
quency (RF) fingerprinting method - or database correlation
method (DCM). It is a network based method which correlates
measured RF fingerprints with RF fingerprints previously stored
in a correlation database (CDB) to estimate the MS location [5].
Those fingerprints contain RF parameters measured by the MS and
reported through the NMR. The CDB might be built from field
measurements or from propagation modeling [6].
The service area - i.e., the region where a LCS based on the
DCM solution is provided - is represented by a digital elevation
model (DEM). The DEM is divided into non-overlapping elements
with a r × r m2 surface each, referred to as pixels. Each pixel has
an associated RF fingerprint stored in the CDB. The correlation
space is the set of pixels whose associated RF fingerprints are
compared, by means of an evaluation function, to the RF fingerprint
measured by the MS which is to be localized. Each pixel within
the correlation space is a candidate solution for the MS location
problem. It is not feasible to include all service area pixels in the
correlation space, as this would cause a very high computational
load, resulting in a long time to produce a position fix. In order to
optimize the search for the MS estimated position, some techniques
can be applied to reduce the correlation space. In this work, this
is done by a genetic algorithm (GA). GA has already been used
together with DCM solutions in [7]. However, we propose an
alternative for the population initialization, restricting the random
distribution of candidate solutions in the first generation to the
predicted best server area of the serving cell (also referred to as
sector). The proposed innovation reduced both the average location
error and the average time to produce a position fix, in comparison
to the original formulation presented in [7].
The remaining of this paper is organized as follows: in Sec-
tion II, the measured and predicted RF fingerprints used in the
MS location are presented, as well as the process of building the
CDB from propagation modeling; in Section III, the correlation
function used to evaluate each position candidate is introduced;
in Section IV, the genetic algorithm used to optimize the MS
location process is presented; in Section V, results for field tests
in a 850 MHz GSM network are analyzed; Section VI brings the
conclusions.
II. RF F INGERPRINTS AND THE CDB
Many DCM solutions in the literature employ CDBs built from
field tests [8] [9]. To keep those CDBs up-to-date, drive tests must
The 7th International Telecommunications Symposium (ITS 2010)
be carried out after any change in radio access network elements,
making this solution impractical for MANs. The detrimental effects
due to the use of out-of-date network parameters in the correlation
process in cellular networks has been demonstrated in [10] and [9].
To prevent those problems, and to allow for a quick and inexpensive
CDB upgrade, a CDB built from propagation modeling has been
used [11] [12]. However, to maximize the accuracy of the DCM
algorithm, additional efforts might be required in initial stages to
tune the propagation models [13].
The propagation model used is this work to build the CDB is
Okumura-Hata [14], which is largely applied to RF planning in
cellular networks. It computes the average propagation loss using:
L p = k1 + k2 log(d) + k3 log(hb ) + k4 L f + k5 log(hb ) log(d) (1)
where d is the distance in meters from the BTS to MS, hb is
the BTS antenna effective height and L f is the diffraction loss
attenuation, calculated by the Epstein-Peterson method [15]. The
MS height is assumed to be 1.5 m. The model coefficients ki
depend on the area morphology and transmission frequency. As the
field test in Section V was carried out in the 869-881 MHz band,
the model was applied at the central frequency of 875 MHz. The
coefficients values are k1 =−12.1, k2 =−44.9, k3 =−5.83, k4 =0.5,
k5 =6.55. All those values are the standard Okumura-Hata values
for urban environment, except k4 , which was empirically defined
by the authors in [6].
The topography of the region( where ) the LCS is to be offered is
represented by a matrix T = ai, j I×J , where ai, j ∈ ℜ+ for any
i ∈ {1, 2, 3, . . . , I} and j ∈ {1, 2, 3, . . . , J}. Each matrix element ai, j
stores the terrain height averaged over a r × r m2 square, where
r is the T matrix planar resolution. The T matrix might also
contain, added to the terrain height, the buildings heights. ⌈ ⌉ If⌈ the⌉
region covers a total surface of L ×W m2 , then T has Lr × Wr
elements. Here, each element of T will be referred to as a pixel.
The received signal strength (RSS) prediction for a single sector,
as shown by Fig. 1, is obtained by the application of Eq. (1) to the
terrain profile between the BTS antenna position and each element
in T . The transmitting antenna characteristics - geographical
coordinates, azimuth, radiation pattern, effective isotropic radiated
power, etc. - are considered in the propagation prediction.
Fig. 1. RSS prediction for a serving antenna.
Each pixel in T has an associated predicted RF fingerprint,
obtained by the superposition of the RSS predictions of all sectors
in the service area. The predicted RF fingerprint of the i-th pixel
is given by the matrix [11]:
⎡ ⎤
IDP,1 RSSP,1
.. ..
ν̂P,i = ⎣ (2)
⎢ ⎥
. . ⎦
IDP,V RSSP,V
The ν̂P,i matrix has V × 2 elements, where V is the number
of sectors in the service area. Each line of ν̂P,i contains a sector
identification (ID) and the correspondent predicted RSS. The lines
are organized in descending order of RSS. The RSS dynamic
range and quantization step vary depending on the wireless access
technology. The CDB is completed after ν̂P has been calculated
for all pixels in T . The predicted RF fingerprints in the correlation
space are compared to the measured RF fingerprint, obtained from
a subset of the parameters listed in the NMR. The measured RF
fingerprint is given by the matrix:
⎡ ⎤
IDM,1 RSSM,1
.. ..
νM = ⎣ (3)
⎢ ⎥
. . ⎦
IDM,N RSSM,N
The νM matrix has N × 2 elements, where N ranges from 1 to
nmax , which is the maximum number of sectors that the MS can
report in the NMR.
III. T HE C ORRELATION F UNCTION
In order to estimate the MS location, it is necessary to compare
the measured RF fingerprint with the predicted RF fingerprints in
the correlation space, which are stored in the CDB. The higher the
similarity or correlation between a measured RF fingerprint and a
predicted RF fingerprint, the higher the probability that the MS is
located at coordinates associated with that predicted RF fingerprint.
This similarity is assessed by means of an evaluation or correlation
function.
Let IP,i be the set of IDs listed in ν̂P,i and let IM be the set of
IDs listed in νM . The set Ci contains the IDs listed both in ν̂P,i and
νM , i.e., Ci = IP,i ∩ IM . Let SP,i and SM be the sets containing
the predicted and measured RSS values, respectively, of the sectors
listed in Ci . The value of the correlation between the measured and
predicted RF fingerprints at the i-th pixel is defined as:
⌊ ⌋
#Ci sM, j − sP,i, j
fi−1 = 2α (nmax − #Ci ) + ∑ (4)
j=1 δ
where δ represents the MS inherent RSS measurement inaccuracy
in dB units [11], nmax is the maximum number of sectors which
can be listed in the NMR, #Ci is the cardinality of Ci , α is the
dynamic range of RSS, sM, j and sP,i, j are the j-th elements of SM
and SP,i , respectively. In GSM networks, nmax = 7, α = 63 dB and
δ = 6 dB [16].
The first term in Eq. (4) is the penalty due to the absence in ν̂P,i
of sectors that are in νM . The second term calculates the absolute
cumulative difference between the elements of SP,i and SM . If all
sectors listed in νM are also in ν̂P,i , i.e., Ci ≡ IP,i ≡ IM , then #Ci =
nmax , and the first term in Eq. (4) is null. In the worst case scenario,
where Ci = ⊘, the first term in Eq. (4) reaches its maximum value
and the second term is null. Regarding the penalty term, consider
the i-th and j-th pixels. For each sector listed in νM and not in
ν̂P,i or ν̂P, j , a 2α value is added to fi or f j , respectively. As α is
equal to the maximum variation of RSS, the first term in Eq. (4)
ensures that if #Ci > #C j , then fi > f j , regardless of the second
term value.
The 7th International Telecommunications Symposium (ITS 2010)
IV. R EDUCING THE C ORRELATION S PACE WITH GA
GA is an adaptive search technique based on the principle of
natural selection and genetic reproduction [17]. Each candidate
solution is an individual, represented by a numeric sequence called
chromosome. When using binary representation, each bit in a
chromosome is referred to as a gene. The set of individuals
at each cycle or generation is called population. The individ-
uals of a population are modified and combined by means of
genetic operators - crossover, mutation, elitism - producing a
new population for the following generation. Crossover mixes
segments of chromosomes of two individuals (parents), producing
two new individuals (crossover children) for the next generation.
Mutation is a random modification of one or more genes of a
chromosome. Elitism is the technique of cloning the best individual
of a generation into the next cycle [18]. The aptitude or fitness
of an individual is assessed by means of an evaluation function.
Better fitted individuals have higher probability of being selected
for reproduction (crossover). The best individual in a population is
the one who achieves the highest value at the evaluation function.
This cycle continues until a stop criterion - maximum number of
generations, fitness of the best individual, processing time, etc. -
has been reached. The best individual of the last generation gives
the problem sub-optimal solution [17].
GA might be used to reduce the correlation space of a DCM
algorithm, optimizing the search for the MS position. The corre-
lation space is the set of pixels whose associated RF fingerprints
are compared to the RF fingerprint measured by the MS. Each
pixel within the correlation space is a candidate solution for the
MS location problem. The set of candidate solutions in a DCM
algorithm, without applying the GA, is the original correlation
space. The set of candidate solutions, after applying the GA, is
the optimized correlation space, which is a subset of the original
correlation space. A GA correlation space reduction factor γ can
be defined as the ratio between the number of elements in the
optimized correlation space and the number of elements in the
original correlation space.
In [7], the original correlation space is the full service area, rep-
resented by T . At the first generation, pa individuals, i.e., pixels,
are randomly selected throughout T . If there are g generations,
the optimized correlation space is a set with pa × g elements, so
γa = (pa × g) /#T , where #T is the cardinality of T .
The probability of a MS being located within the predicted best
server area of its serving sector is higher than in any other pixels
in the service area. Therefore, the original correlation space is
restricted to the predicted best server
{ area of the serving } sector,
defined by the set of pixels X = i ∈ T ∣ IDP,i,1 = IDM,1 , where
IDP,i,1 is the predicted best server ID at the i-th pixel and IDM,1
is the measured best server ID, reported in the NMR. So, when
initializing the first generation population, instead of randomly
selecting individuals throughout the whole service area, the in-
dividuals should be randomly selected among the pixels within
X . In such conditions, if the population size is pb individuals
and there are g generations, the correlation space reduction factor
is γb = (pb × g) /#X . Two assumptions can be made regarding
the proposed innovation in the initialization of the first generation
population:
1) X ⊂ T , so #X ≤ #T . In practice, #X << #T , because
the whole service area is much wider than a single sector
best server area. As a result, for γa = γb , pb < pa , which
means that, in comparison to [7], less individuals will be
required per generation to find the sub-optimal solution;
2) As the probability of a MS being located within the pixels
in X is higher than in any other subset of T , the average
fitness of the first generation population will be higher - i.e.,
in average, the first generation individuals will be closer to
the real MS location - which means that, in comparison
to [7], less generations will be required to find the sub-
optimal solution;
Therefore, it is expected that the use of the proposed innovation
in the initialization of the first generation population, will result
in higher accuracy and lower time to produce a position fix. This
will be verified in Section V.
The proposed GA steps can be summarized as:
1) Initialize first generation population, randomly selecting
individuals - i.e., pixels or candidate locations - within the
original correlation space, which can be the full service area,
represented by matrix T , or the predicted best server area
of the serving sector, represented by matrix X ;
2) Evaluate the fitness of each individual in current population,
using the evaluation function in Eq. (4);
3) Create chromosomes, converting the individuals coordinates
to binary format;
4) Apply genetic operators - crossover, mutation and elitism -
to create a new generation;
5) Convert chromosomes to integer format;
6) If stop criterion has been met, provides MS location, given
by the coordinates of the fittest individual; otherwise, returns
to step 2.
2
⌉ with L × W m , then T is a
If the service area⌈ is⌉ a ⌈square
L W
square matrix with r × r elements. If the service area is
geographically represented by means of a rectangular projection,
like Universal Transverse Mercator (UTM) [19], and the UTM
coordinates of one pixel are known, then it is possible to obtain the
UTM (x, y) coordinates of any other element in T . The length of
each chromosome will be the number of bits required to identify
the position
⌈( of a⌈ pixel - i.e., ⌉)⌉line and column in T - and is
⌈ its
given by log2 Lr + log2 Wr
⌉
.
V. F IELD T ESTS IN A 850 MH Z GSM N ETWORK
A. Drive Test Scenario
Field tests were performed in a 850 MHz GSM network in
the downtown area of Rio de Janeiro. The test area characteristics
are summarized in Tab. I. The test set was composed of a GSM
phone and a GPS receiver, both connected to a laptop placed
inside a moving vehicle. The MS was in active mode and for
each transmitted NMR the current location was calculated by the
GPS receiver. Every NMR and GPS measurement were recorded
for further processing. A total number of 4501 NMRs has been
collected. The large number of samples ensures a reasonable
confidence for statistical analysis. The GPS location was assumed
to be the reference position, so, for each NMR and each location
method, the positioning error is the Euclidean distance between the
GPS position and the location provided by the respective method.
B. GA Parameters
The parameters used in the genetic algorithm are summarized
in Tab. II. Parameters gmax , ε and φ define the stop criterion. The
GA stops when one of the two conditions occurs: i) the maximum
number of generation gmax is reached; ii) the fitness of the best
individual during φ consecutive generations does not improve by a
The 7th International Telecommunications Symposium (ITS 2010)

value higher than ε . The second condition is an improvement of the

1

0.9
stop criterion based only on the maximum number of generations: Proposed (II)
Original (I)
if the aptitude of the best individual reaches a steady state, it might 0.8

mean that the algorithm has reached a local maximum and therefore 0.7

there is no use in creating new generations [20]. Parameter γ is the 0.6

Percentle
correlation space reduction factor, defined in Section IV. 0.5

0.4
TABLE I
0.3
C HARACTERISTICS OF THE TEST AREA
0.2

0.1
Characteristic Value
Morphology Dense Urban 0
0 100 200 300 400 500 600
Area 4.84 km2 Error (meters)

Cell Density 24 cells/km2 Fig. 2. CDF of MS location error in meters.

DEM planar resolution (r) 10 m
DEM with building heights? Yes −3
x 10
8
Proposed (II)
Original (I)
7

TABLE II
6
S UMMARY OF GA PARAMETERS
5

Fitness
Parameter Value
Crossover ratio 60% 4

Selection type Roulette

3
Mutation ratio 1%
Elitism Yes 2
Chromosome length 16 genes (bits)
gmax 20 generations 1
ε
0 500 1000 1500 2000 2500 3000 3500 4000 4500
0.00001 NMR Number
φ 5 generations Fig. 3. Average fitness of first generation per NMR.
γ 3%

C. Experimental Results
Fig. 2 shows the cumulative distribution function (CDF) of the
MS location error achieved by the DCM using the original (Method
I) and proposed (Method II) initialization methods for the first
generation population. Both have used the parameters and stop
criterion described in Section V-B. Method I has achieved location
errors of 197 and 426 meters, for the 50-th and 90-th percentiles.
Method II has achieved location errors of 156 and 362 meters for
the same percentiles, which represents an error reduction of 20%
and 15%, respectively.
Fig. 3 shows the average fitness of the first population for each
positioning fix, i.e per collected NMR. As expected, Method II
achieved a higher average fitness in the first generation population,
what results in less generations to reach the suboptimal solution:
Method I requires in average 9 generations, while Method II
requires only 7 generations - a 23% reduction.
The reduction in the average number of generations, coupled
with the fact that the population size in Method II is much smaller
than in Method I, results in a significant reduction in the time
- i.e., the current MS location estimate is given by the arithmetic
mean of the previous 20 estimated positions - to the results obtained
by Method II, the 50-th and 98-th location error percentiles become
119 and 342 meters. This precision is slightly better than the one
achieved in the third route in [22] - 125 and 349 meters for the
same percentiles - where a GA has been used for MS tracking.
However, in [22] the test has been carried out in a 1800 MHz
GSM network with 41 cells/km2 . This cell density is considerably
higher than in the test area described in Tab. I. Therefore, it
is expected that the DCM algorithm presented here, using GA
with the proposed first generation population initialization, would
achieve an even higher precision if applied to the same environment
described in [22].
0.5
Proposed (II)
Original (I)
0.45
0.4

required to produce a position fix, as shown in Fig. 4. To allow 0.35

for a fair comparison, both methods were tested using the same 0.3
Seconds

hardware and software. The average time required to produce a 0.25

position fix using Method II was 91% lower than in Method I. 0.2

This makes Method II a more suitable option for real time MS 0.15
location applications, like MS tracking. 0.1
In MS tracking applications, a moving average filter might
0.05
be helpful in reducing the location error, by filtering out abrupt
0
variations in location estimations between adjacent position fixes 0 500 1000 1500 2000 2500 3000 3500 4000 4500
NMR Number
along a route [21]. Applying a moving average filter with length 20 Fig. 4. Time in seconds to produce a position fix per NMR.
The 7th International Telecommunications Symposium (ITS 2010)

VI. C ONCLUSION [14] M. Hata, “Empirical Formula for Propagation Loss in Land Mobile
Radio Services,” IEEE Transactions on Vehicular Technology, vol. 29,
This paper proposes an alternative for the initialization of the no. 3, pp. 317–325, August 1980.
first generation population of a genetic algorithm, used to reduce [15] M. D. Yacoub, Foundations of Mobile Radio Engineering. CRC
the correlation space in a RF fingerprinting location method. The Press, 1993.
proposed alternative consists in restraining the randomly selected [16] European Telecommunications Standard Institute, ETSI TS 100911
v6.2.0 (1998-07) - Digital cellular telecommunications system (Phase
first population individuals to the predicted best server area of
2+); Radio subsystem link control (GSM 05.08 version 6.2.0 Release
the serving sector, instead of selecting individuals throughout the 1997), 1998.
whole service area. The proposed innovation produces smaller and [17] P. M. Pardalos, L. Pitsoulis, T. Mavridou, and M. G. C. Resende,
better fitted initial populations, which results in location accuracy “Parallel Search for Combinatorial Optimization: Genetic Algorithms,
improvement and in significant reduction of the average time to Simulated Annealing, Tabu Search and GRASP,” in Proceedings of
produce a position fix. The method has been tested in a 850 MHz the 2nd International Workshop on Parallel Algorithms for Irregularly
Structured Problems, Lyon, France, September 1995, pp. 317–332.
GSM network, but its application is extensible to different types of
[18] D. Goldberg, Genetic Algorithms in Search, Optimization and Ma-
wireless metropolitan area networks. The achieved precision shows chine Learning. Addison-Wesley, 1989.
that the DCM+GA proposed location method is suitable for MS [19] International Telecommunications Union, “Digital Topographical
positioning in urban environments. Database for Propagation Studies,” Recommendation ITU-R P.1058-1,
1997.
R EFERENCES [20] C. Houck, J. Joines, and M. Kay, “A Genetic Algorithm for Function
Optimization: A Matlab Implementation,” NCSU-IE TR 95-09, 1995.
[1] Federal Communications Commission, “Guidelines for Testing and [21] X. Shen, J. W. Mark, and J. Ye, “Mobile Location Estimation in
Verifying the Accuracy of Wireless E911 Location Systems,” OET Cellular Networks Using Fuzzy Logic,” in Proceedings of IEEE 52nd
BULLETIN No.71, 2000. Vehicular Technology Conference, Boston, USA, September 2000, pp.
[2] European Commission, “Commission Recommendation 2108–2114.
2003/558/EC,” Official Journal of the European Union, 2003. [22] R. T. J. D. B. Lin and H. P. Lin, “Mobile Location Estimation
and Tracking for GSM Systems,” in Proceedings of IEEE 15th
[3] M. A. Spirito, S. Pykko, and O. Knuuttila, “Experimental Performance
International Symposium on Personal, Indoor and Mobile Radio
of Methods to Estimate the Location of Legacy Handsets in GSM,” in
Communications, Barcelona, Spain, September 2004.
Proceedings of IEEE 54th Vehicular Technology Conference, Atlantic
City, USA, October 2001, pp. 2716–2720.
[4] M. I. Silventoinen and T. Rantalainen, “Mobile Station Emergency
Locating in GSM,” in Proceedings of IEEE International Conference
on Personal Wireless Communications, New Delhi, India, February
1996, pp. 232–238.
[5] H. Laitinen, J. Lahteenmaki, and T. Nordstrom, “Database Correlation
Method for GSM Location,” in Proceedings of IEEE 53rd Vehicular
Technology Conference, Rhodes, Greece, May 2001, pp. 2504–2508.
[6] R. S. Campos and L. Lovisolo, “A Fast Database Correlation Al-
gorithm for Localization of Wireless Network Mobile Nodes using
Coverage Prediction and Round Trip Delay,” in Proceedings of IEEE
69th Vehicular Technology Conference, Barcelona, Spain, April 2009.
[7] M. J. Magro and C. J. Debono, “A Genetic Algorithm Approach to
User Location Estimation in UMTS Networks,” in EUROCON - The
International Conference on Computer as a Tool, Warsaw, Poland,
September 2007, pp. 1136–1139.
[8] B. Li, Y. Wang, H. K. Lee, A. Dempster, and C. Rizos, “A New
Method for Yielding a Database of Location Fingerprints in WLAN,”
IEE Proceedings Communications, vol. 152, no. 5, pp. 580–586,
October 2005.
[9] M. A. Hallak, M. Safadi, and R. Kouatly, “Mobile Positioning
Technique using Signal Strength Measurement method with the aid
of Passive Mobile Listener Grid,” in Proceedings of 2nd International
Conference on Information and Communication Technologies: from
Theory to Applications, Damascus, Syria, April 2006, pp. 105–110.
[10] M. Chen, T. Sohn, D. Chmelev et al., “Practical Metropolitan-Scale
Positioning for GSM Phones,” in Proceedings of 8th International
Conference on Ubiquitous Computing, Newport Beach, USA, Septem-
ber 2006, pp. 225–242.
[11] R. S. Campos and L. Lovisolo, “Location Methods for Legacy GSM
Handsets using Coverage Prediction,” in Proceedings of IEEE 9th
Workshop on Signal Processing Advances in Wireless Communica-
tions, Recife, Brazil, July 2008, pp. 21–25.
[12] D. Zimmermann, J. Baumann, A. Layh, F. Landstorfer, R. Hoppe, and
G. Wlfle, “Database Correlation for Positioning of Mobile Terminals
in Cellular Networks using Wave Propagation Models,” in Proceed-
ings of IEEE 60th Vehicular Technology Conference, Los Angeles,
USA, September 2004, pp. 4682–4686.
[13] J. Zhu and G. D. Durgin, “Indoor/Outdoor Location of Cellular
Handsets Based on Received Signal Strength,” Eletronic Letters,
vol. 41, no. 1, pp. 24–26, January 2005.